SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2410.18538 · ICLR · 2025

SMITE: Segment Me In TimE

Saeid Taghanaki, Andrea Tagliasacchi, Sauradip Nag, Ghassan Hamarneh, Amirhossein Alimohammadi, Ali Amiri

arXiv · PDF · Open in the Atlas

Code that ran

14 of 21 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 4 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 10 failures of our instrument; 7 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2410.18538.json · badge: /api/ran/2410.18538.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
alimohammadiamirhossein/smite reimplementation (classifier label)
classifier basis: classifier_2025
14 ran · 7 no recorded run
FunctionStatusWhere it lives
_calculate_pixel_weights Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("7227bc7182f55a3b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 1, dtype family float, shape (5,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockPipeline(nn.Module):
    def __init__(self, num_parts=5, device='cpu'):
        super().__init__()
        self.num_parts = num_parts
        self.device = device

pipeline = MockPipeline(num_parts=5, device='cpu')

was_tracked = torch.randint(0, 2, (2, 4), dtype=torch.float32)

ARGS = {
    "pipeline": pipeline,
    "was_tracked": was_tracked
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 1, dtype family float, shape (5,) — shape (5,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_compute_loss Ran
ran, honoured the contract we drafted
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("01aa7ec519d76e3b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockPipeline(nn.Module):
    def __init__(self, num_parts=5, device='cpu'):
        super().__init__()
        self.num_parts = num_parts
        self.device = device

# Create small tensors with distinct dimensions
# was_dst: (batch=2, frames=3, channels=4, height=5, width=6)
was_dst = torch.randn(2, 3, 4, 5, 6)

# was_refrence: (batch=2, frames=3, channels=4, height=5, width=6)
was_refrence = torch.randn(2, 3, 4, 5, 6)

# was_tracked: (batch=2, frames=3, num_parts=5, height=5, width=6)
# This represents class logits for each part
was_tracked = torch.randn(2, 3, 5, 5, 6)

# Create pipeline mock
pipeline = MockPipeline(num_parts=5, device='cpu')

# Create latents_dict
latents_dict = {
    'was_refrence': was_refrence,
    'was_tracked': was_tracked,
}

# Create kwargs with various options
kwargs = {
    'regularization_weight': 0.5,
    'track_weight': 0.3,
    'dct_threshold': 0.2,
    'i': 0,
}

ARGS = {
    'pipeline': pipeline,
    'latents_dict': latents_dict,
    'was_dst': was_dst,
    'kwargs': kwargs,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch, torch_dct — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_compute_refrence_loss Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("a40ca276c0314b43")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create small tensors for feat_pred and feat_gt
# Expected shape after unsqueeze(0): (b, f, c, h, w)
# feat_pred and feat_gt should be (f, c, h, w) before unsqueeze
feat_frames = 2
feat_channels = 4
feat_height = 3
feat_width = 5

was_dst = torch.randn(feat_frames, feat_channels, feat_height, feat_width)
was_refrence = torch.randn(feat_frames, feat_channels, feat_height, feat_width)

# Create kwargs dict
kwargs = {
    'i': 0,
    'dct_threshold': 0.2
}

ARGS = {
    'was_dst': was_dst,
    'was_refrence': was_refrence,
    'kwargs': kwargs
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch, torch_dct — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_compute_tracking_loss Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("be631dd6f50039f2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockPipeline(nn.Module):
    def __init__(self, num_parts=5, device='cpu'):
        super().__init__()
        self.num_parts = num_parts
        self.device = device

# Create pipeline
pipeline = MockPipeline(num_parts=5, device='cpu')

# Create tensors
# was_dst: float tensor of shape (batch_size, num_classes) for cross_entropy input
# was_tracked: tensor of shape (batch_size, num_classes) with class logits/scores
batch_size = 2
num_classes = 5

was_dst = torch.randn(batch_size, num_classes)
was_tracked = torch.randn(batch_size, num_classes)

ARGS = {
    'was_dst': was_dst,
    'was_tracked': was_tracked,
    'pipeline': pipeline,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_process_timestep Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("76eb30418deb0324")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from typing import Dict, List, Optional, Union

torch.manual_seed(0)

# Build minimal mock scheduler
class MockScheduler:
    def scale_model_input(self, latent_model_input, t):
        return latent_model_input
    
    def step(self, noise_pred, t, latents, **kwargs):
        class StepOutput:
            def __init__(self):
                self.prev_sample = latents - 0.01 * noise_pred
        return StepOutput()

# Build minimal mock controlnet
class MockControlNet(nn.Module):
    def forward(self, control_model_input, timestep, encoder_hidden_states, controlnet_cond, conditioning_scale, return_dict=False):
        batch_size = control_model_input.shape[0]
        down_samples = [torch.randn(batch_size, 4, 8, 8) for _ in range(4)]
        mid_sample = torch.randn(batch_size, 8, 4, 4)
        return (down_samples, mid_sample)

# Build minimal mock unet
class MockUNet(nn.Module):
    def forward(self, latent_model_input, t, encoder_hidden_states, mid_block_additional_residual=None, down_block_additional_residuals=None, cross_attention_kwargs=None, **kwargs):
        class UNetOutput:
            def __init__(self):
                self.sample = torch.randn(latent_model_input.shape[0], 4, 8, 8)
        return UNetOutput()

# Build minimal mock pipeline with required methods
class MockPipeline(nn.Module):
    def __init__(self):
        super().__init__()
        self.scheduler = MockScheduler()
        self.controlnet = MockControlNet()
        self.unet = MockUNet()
        self.attention_maps = {}
    
    def get_attention_map(self, attention_maps, output_size, token_ids, average_layers, apply_softmax, video_length, guidance_scale, **kwargs):
        cross_attn = torch.randn(2, 10, 64)
        self_attn = torch.randn(2, 64, 64)
        return (None, cross_attn, self_attn)
    
    def calculate_was_mask(self, cross_attention_maps, self_attention_maps, mask_patch_size, use_cross_attention_only=False):
        return torch.randn(2, 512, 512)

# Construct all arguments
pipeline = MockPipeline()

prompt_embeds = torch.randn(2, 10, 768)
latents = torch.randn(2, 4, 8, 8)
depth_input = torch.randn(2, 3, 512, 512)
t = 500
attention_output_size = 64
token_ids = [0, 1, 2, 3, 4]
average_layers = False
apply_softmax = True
video_length = 16
guidance_scale = 7.5
cross_attention_kwargs = {}
controlnet_keep = torch.ones(1000)
controlnet_conditioning_scale = 1.0
extra_step_kwargs = {}

ARGS = {
    'pipeline': pipeline,
    'prompt_embeds': prompt_embeds,
    'latents': latents,
    'depth_input': depth_input,
    't': t,
    'attention_output_size': attention_output_size,
    'token_ids': token_ids,
    'average_layers': average_layers,
    'apply_softmax': apply_softmax,
    'video_length': video_length,
    'guidance_scale': guidance_scale,
    'cross_attention_kwargs': cross_attention_kwargs,
    'controlnet_keep': controlnet_keep,
    'controlnet_conditioning_scale': controlnet_conditioning_scale,
    'extra_step_kwargs': extra_step_kwargs,
    'return_latent': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_update_moments Ran
ran, honoured the contract we drafted
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("3c743feea5734a20")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Analyze the function:
# - grad_cond: used in arithmetic operations (multiplication, exponentiation)
#   Should be a scalar or tensor
# - kwargs with keys: 'm1', 'm2', 'iter_'
# - All operations are element-wise arithmetic

# grad_cond should be a small tensor or scalar
grad_cond = torch.tensor(0.5)

ARGS = {
    'grad_cond': grad_cond,
    'm1': torch.tensor(0.1),
    'm2': torch.tensor(0.05),
    'iter_': 0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calculate_controlnet_features Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("0cb10cbe5854515a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockControlNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def forward(self, sample, timestep, encoder_hidden_states=None, controlnet_cond=None, conditioning_scale=None, return_dict=False):
        # Return down_block_res_samples (list of tensors) and mid_block_res_sample (tensor)
        down_block_res_samples = [torch.randn(2, 4), torch.randn(2, 4)]
        mid_block_res_sample = torch.randn(2, 4)
        return (down_block_res_samples, mid_block_res_sample)

class MockPipeline:
    def __init__(self):
        self.controlnet = MockControlNet()

ARGS = {
    "pipeline": MockPipeline(),
    "depth": torch.randn(2, 4),
    "latent_model_input": torch.randn(2, 4),
    "prompt_embeds": torch.randn(2, 4),
    "timestep": 42,
    "controlnet_keep": [1.0] * 1000,
    "controlnet_conditioning_scale": 1.0,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calculate_new_was_value Ran
ran, honoured the contract we drafted
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("b8df771e20c60389")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a small tensor of shape (2, 4) with float values
# This will be used as features for 2 samples with 4 dimensions
feats = torch.randn(2, 4)

ARGS = {
    "feats": feats
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
colorize_depth_maps Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("b5c8182dd8640ec2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 3, 4, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
import matplotlib

torch.manual_seed(0)

# Create a small depth map as torch tensor
depth_map_tensor = torch.randn(2, 4, 5)  # batch=2, height=4, width=5

# Create a valid mask
valid_mask = torch.ones(2, 4, 5, dtype=torch.bool)

# Define min and max depth values
min_depth = 0.0
max_depth = 1.0

# Define colormap
cmap = 'Spectral'

ARGS = {
    'depth_map': depth_map_tensor,
    'min_depth': min_depth,
    'max_depth': max_depth,
    'cmap': cmap,
    'valid_mask': valid_mask
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 3, 4, 5) — shape (2, 3, 4, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
create_slices Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("4cb5f2bdcabf144d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "length": 100,
    "slice_size": 30,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
dct_loss Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("fe6f43be61a8c67d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn.functional as F
from einops import rearrange
import torch_dct

torch.manual_seed(0)

# Create tiny feature tensors
# feat_pred and feat_gt should have shape (F, C, H, W) before unsqueeze
# After unsqueeze(0): (1, F, C, H, W)
# After rearrange to 'b c f h w': (1, C, F, H, W)
F_frames = 2
C_channels = 3
H_height = 4
W_width = 5

feat_pred = torch.randn(F_frames, C_channels, H_height, W_width)
feat_gt = torch.randn(F_frames, C_channels, H_height, W_width)

ARGS = {
    "feat_pred": feat_pred,
    "feat_gt": feat_gt,
    "low_pass": True,
    "threshold": 0.2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch, torch_dct — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
dct_low_pass_filter Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("d5f0c2ad9c03b52a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 5, dtype family float, shape (2, 3, 2, 4, 6) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a 2D tensor of DCT coefficients
# The function expects shape with at least 5 dimensions based on indexing: [:, :, :, :cutoff_x, :cutoff_y]
# So we need a 5D tensor: (batch, channels, depth, height, width)
dct_coefficients = torch.randn(2, 3, 2, 4, 6)

ARGS = {
    "dct_coefficients": dct_coefficients,
    "percentage": 0.3
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 5, dtype family float, shape (2, 3, 2, 4, 6) — shape (2, 3, 2, 4, 6) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
determine_window Ran
ran, honoured the contract we drafted
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("ad464697812cba58")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# The function determine_window takes:
# - index: an integer index position
# - length: the total length of a sequence
# - window_size: the size of the window (default 7, must be odd)

# Reading the function body:
# - index is compared with half_window and length - half_window (integer comparisons)
# - length is used in comparisons and arithmetic
# - window_size is checked for oddness (% 2), divided by 2, and used in slicing tuples
# All parameters are simple integers

ARGS = {
    'index': 5,
    'length': 20,
    'window_size': 7
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family int — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
gather_features Ran
ran; our contract draft was wrong, not the code
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("bec020dc9147bb3b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Build tensors
batch_size = 2
num_features = 4
height = 3
width = 5

# was: indexed as was[j, :, coords[1], coords[0]] -> [batch, features, height, width]
was = torch.randn(batch_size, num_features, height, width)

# pred_track_indice: indexed as pred_track_indice[j, k] -> returns coords [x, y]
# Shape [batch, num_tracks, 2] where last dim is [x, y]
pred_track_indice = torch.randint(0, width, (batch_size, 6, 2))
pred_track_indice[..., 1] = torch.randint(0, height, (batch_size, 6))

# pred_visibility: indexed as pred_visibility[j, k] -> scalar
# Shape [batch, num_tracks]
pred_visibility = torch.rand(batch_size, 6)

# Scalar parameters
i = 0
k = 1
start = 0
end = 2

ARGS = {
    "was": was,
    "pred_track_indice": pred_track_indice,
    "pred_visibility": pred_visibility,
    "i": i,
    "k": k,
    "start": start,
    "end": end,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_optimize_a_slice Not yet run
unverified (harvested, no recorded run)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("18387f7bb11f3441")
How it ran

no invocation recorded.

_update_latent_with_energy_function Not yet run
unverified (harvested, no recorded run)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("996cce05a83e5dc3")
How it ran

no invocation recorded.

_update_latent_with_gradient Not yet run
unverified (harvested, no recorded run)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("8de285a460a7632e")
How it ran

no invocation recorded.

_visualization_during_optimization Not yet run
unverified (harvested, no recorded run)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("0997043fb02a185a")
How it ran

no invocation recorded.

apply_voting Not yet run
unverified (harvested, no recorded run)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("a355f21b2457d26e")
How it ran

no invocation recorded.

optimize_latents Not yet run
unverified (harvested, no recorded run)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("2942ddefa7ca0a88")
How it ran

no invocation recorded.

visualize_attention_maps Not yet run
unverified (harvested, no recorded run)
alimohammadiamirhossein/smite/src/latent_optimization.py
code served (permissive licence) · get_code("ef527e7f7a4c8410")
How it ran

no invocation recorded.

Also present in other papers' code

Of 21 samples lifted from this paper's linked repositories, 21 carry a function identity and 0 do not. Of the 21 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 20 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
colorize_depth_maps
brick of this paper's assembly: helper
1 other paper
1 dated earlier
2312.02145 (2023-12)
Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Figure 1: SMITE. Using only one or few segmentation references with fine granularity (left), our method learns to segment different unseen videos respecting the segmentation references.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2410.18538")
get_code_for_paper("2410.18538")
have("2410.18538")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2410.18538.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2410.18538.svg)](https://syntology.ai/paper/2410.18538)

Connect an agent — have() is free.